A high-profile AI school has had its headline claim tested against its own data. The New York Times published an investigation on October 23, 2026 by Dana Goldstein into Alpha School, the network that markets AI-powered learning and "2x" student growth. According to excerpts quoted by education researcher Benjamin Riley in a same-day reaction post, Goldstein obtained internal learning data and found that "only 42 percent progressed twice as fast in both math and reading, meeting the '2x' growth standard featured in the company's marketing."
We could not open the Times article, which is paywalled, so this post does not reproduce its reporting. It sets out the claim, the reported data, what Alpha has said publicly, and how to read all of it. Treat the Times figures as second-hand until you read the original.
TL;DR: claim versus reported data
| Question | Answer |
|---|---|
| What is the claim? | AI-powered "2 Hour Learning" produces about 2x the learning growth of typical schooling |
| Who reported the data? | The New York Times (Dana Goldstein), per excerpts quoted by Benjamin Riley |
| Headline figure | 42% progressed at least twice as fast in both math and reading |
| Lower performers | About a fifth of private-school students were at or below the 50th percentile in math last fall; 37% of that group met 2x growth by winter |
| What we verified | The Riley post and Alpha's own earlier blog post; not the Times article |
| What is unverified | Exact definitions, sample sizes and the full data set |
| Independent audit? | None found |
What Alpha School claims
Alpha's public pitch is that students spend roughly two hours a day on AI-driven academics using software, with the rest of the day on projects and life skills, and that this produces elite outcomes. Its co-founder MacKenzie Price wrote in a February 2026 blog post that the school's principal fed its Fall 2025 to Winter 2026 NWEA MAP dataset into Gemini and Claude and asked how Alpha compares to elite private schools.
The post reproduces the Gemini 3 Pro output, which describes the results as "statistically exceptional," places the district beyond the 99th percentile and compares it to schools such as Phillips Exeter. It cites mean scores such as a kindergarten math RIT of 177.2 and a grade 11 math RIT of 278.3. Those are claims produced by asking an AI model to interpret the school's own numbers, not an external evaluation, and an LLM's praise of a dataset is not verification. That distinction is central to the story, because the same data now looks different when a journalist examines the growth distribution rather than the averages.
What the Times reportedly found
Riley, an education critic who has written critical reports on Alpha since September, quotes two passages. First, the 42 percent figure for meeting the "2x" standard in both subjects. Second, that Alpha "had an especially difficult time meeting that mark for its lower-performing math students": about a fifth of private-school students scored at or below the 50th percentile nationally in math last fall, and only 37 percent of that group met the promised 2x growth by winter.
Two points for careful reading:
- "Both math and reading" is a strict bar. A student who doubles growth in math but not reading is not counted in the 42 percent. Alpha might argue for per-subject rates, and the Times article likely gives them; we could not see them.
- Average versus distribution. High average percentile scores can coexist with a minority of students not reaching 2x growth, especially if a school enrolls many students who started high, where growth is harder to measure because of ceiling effects on the test.
Riley himself is an interested party. He is a vocal critic who says Alpha is not talking to him, and he calls parts of the Times story underwhelming ("kind of a big nothingburger," quoting a source) while still praising its core finding. His claims about Alpha's student-social-media requirements and other practices are his own reporting and we have not verified them.
What Alpha has said, and what it has not
We did not find a response from Alpha to the Times figures in the material we could open. If Alpha responds, the relevant questions will be about the definition of 2x growth, the comparison cohort and which students are included.
Alpha's own February post rests on its internal NWEA MAP dataset, and we found no independent review of it. The Times data, as described, appears to be the first time an outside party has published figures drawn from the network's records, which is why this story matters beyond a single school.
How the scrutiny built up in 2026
The Times story follows a short sequence of public material, each covering a different part of the claim:
- February 17, 2026: Alpha publishes its own mid-year "report card," built on an AI model's reading of its MAP data, as described above.
- Early September 2026: Riley says he published the first of his special reports, focused on an Alpha boot camp, and says more are coming.
- October 23, 2026: The Times publishes growth figures from internal data, including the 42 percent share.
The step from the school's own narrative toward outside evidence is common with AI products: first a vendor-run analysis, then a data-driven check of the headline metric. It is why we label stories like this "claimed versus verified" rather than picking a side. The claim is clear, the first outside numbers exist, and the full evidence is still being assembled.
How to judge a "2x learning" claim
If a school, an edtech vendor or an AI tutor tells you it doubles learning, ask:
- Doubles relative to what? MAP publishes typical growth norms by starting score and grade. Doubling needs a baseline.
- What share of students, not what average? The Times figure is a share (42 percent), which is more informative than a mean.
- Both subjects or either? Requirements change the number a lot.
- Who analyzed it, and with what? An AI model summarizing a school's own data is not an audit. Look for third-party analysis or a published methodology.
- What about students who start low? The reported 37 percent figure for lower-performing math students is the kind of subgroup number that headline claims tend to leave out.
- What else was changed? Selective admission, tuition, family income and time on task all affect results, and they are separate from whether the software helps.
These questions apply equally to the broader wave of AI products that publish their own outcome figures. We have covered the same pattern in the AI verification tax surveys, in Pine's SaaS benchmark claims, and in the Paper Instruments Ultramarine claim. When the party selling a result is also the one measuring it, independent data changes the picture.
Why it matters for AI in education
Alpha is a prominent example of an AI-centered school model, and if districts consider adopting similar programs, a private-school marketing claim becomes a policy question: would they be buying a model whose headline number does not describe most students? AI tutoring may still help many students; the Times figures, as quoted, do not show that it fails, they show that a specific, widely repeated number did not describe 58 percent of the students measured, on a strict definition.
For builders of AI learning products the lesson is practical. Publish the distribution, define the metric, let outsiders reproduce it, and expect journalists to ask for the data. For parents and districts: ask for share-of-students figures by starting level, and read the original reporting in full.
A worked example of why the share matters
Consider a hypothetical school of 100 students where the average student grows 2.1 times as fast as the national norm. That sounds like a clear win. But if 30 students grow at four times the norm and 70 grow at 1.2 times the norm, the average is pulled up by a minority, and only 30 percent of students actually met a 2x bar. The average and the share answer different questions. A parent deciding whether a school will work for their child cares about the share, and especially about the share among children who look like theirs. This is why the reported 37 percent figure for lower-performing math students is more informative to many families than a school-wide mean, and why the strict "both math and reading" definition reported by the Times changes the headline so much.
The same logic applies to any outcome claim from an AI product: a vendor can truthfully report a mean while a large group sees little benefit. Asking for the distribution is not hostility to the product; it is the minimum needed to interpret the number.
What to watch next
- The Times article itself, for exact definitions, sample sizes and Alpha's response.
- Any statement from NWEA about how its MAP growth data is being used in marketing.
- Whether districts piloting Alpha-style programs request independent evaluation as a condition of contracts.
- Further reporting from Riley and others, which should be read as advocacy as well as journalism.
Details are accurate as of October 23, 2026, based on a reaction post quoting the Times investigation and Alpha's own blog. We did not have access to the Times article.